I’m more active on Twitter than LW/AF these days: https://twitter.com/DavidSKrueger
https://www.davidscottkrueger.com/
David Scott Krueger (formerly: capybaralet)
Karma: 2,176
First, RE the role of “solving alignment” in this discussion, I just want to note that:
1) I disagree that alignment solves gradual disempowerment problems.
2) Even if it would that does not imply that gradual disempowerment problems aren’t important (since we can’t assume alignment will be solved).
3) I’m not sure what you mean by “alignment is solved”; I’m taking it to mean “AI systems can be trivially intent aligned”. Such a system may still say things like “Well, I can build you a successor that I think has only a 90% chance of being aligned, but will make you win (e.g. survive) if it is aligned. Is that what you want?” and people can respond with “yes”—this is the sort of thing that probably still happens IMO.
4) Alternatively, you might say we’re in the “alignment basin”—I’m not sure what that means, precisely, but I would operationalize it as something like “the AI system is playing a roughly optimal CIRL game”. It’s unclear how good of performance that can yield in practice (e.g. it can’t actually be optimal due to compute limitations), but I suspect it still leaves significant room for fuck-ups.
5) I’m more interested in the case where alignment is not “perfectly” “solved”, and so there are simply clear and obvious opportunities to trade-off safety and performance; I think this is much more realistic to consider.
6) I expect such trade-off opportunities to persist when it comes to assurance (even if alignment is solved), since I expect high-quality assurance to be extremely costly. And it is irresponsible (because it’s subjectively risky) to trust a perfectly aligned AI system absent strong assurances. But of course, people who are willing to YOLO it and just say “seems aligned, let’s ship” will win. This is also part of the problem...
My main response, at a high level:
Consider a simple model:We have 2 human/AI teams in competition with each other, A and B.
A and B both start out with the humans in charge, and then decide whether the humans should stay in charge for the next week.
Whichever group has more power at the end of the week survives.
The humans in A ask their AI to make A as powerful as possible at the end of the week.
The humans in B ask their AI to make B as powerful as possible at the end of the week, subject to the constraint that the humans in B are sure to stay in charge.
I predict that group A survives, but the humans are no longer in power. I think this illustrates the basic dynamic. EtA: Do you understand what I’m getting at? Can you explain what you think it wrong with thinking of it this way?
Responding to some particular points below:
Sure, but these things don’t result in non-human entities obtaining power right?
Yes, they do; they result in beaurocracies and automated decision-making systems obtaining power. People were already having to implement and interact with stupid automated decision-making systems before AI came along.
Like usually these are somewhat negative sum, but mostly just involve inefficient transfer of power. I don’t see why these mechanisms would on net transfer power from human control of resources to some other control of resources in the long run. To consider the most extreme case, why would these mechanisms result in humans or human appointed successors not having control of what compute is doing in the long run?
My main claim was not that these are mechanisms of human disempowerment (although I think they are), but rather that they are indicators of the overall low level of functionality of the world.
I think we disagree about:
1) The level of “functionality” of the current world/institutions.
2) How strong and decisive competitive pressures are and will be in determining outcomes.
I view the world today as highly dysfunctional in many ways: corruption, coordination failures, preference falsification, coercion, inequality, etc. are rampant. This state of affairs both causes many bad outcomes and many aspects are self-reinforcing. I don’t expect AI to fix these problems; I expect it to exacerbate them.
I do believe it has the potential to fix them, however, I think the use of AI for such pro-social ends is not going to be sufficiently incentivized, especially on short time-scales (e.g. a few years), and we will instead see a race-to-the-bottom that encourages highly reckless, negligent, short-sighted, selfish decisions around AI development, deployment, and use. The current AI arms race is a great example—Companies and nations all view it as more important that they be the ones to develop ASI than to do it carefully or put effort into cooperation/coordination.
Given these views:
1) Asking AI for advice instead of letting it take decisions directly seems unrealistically uncompetitive. When we can plausibly simulate human meetings in seconds it will be organizational suicide to take hours-to-weeks to let the humans make an informed and thoughtful decision.
2) The idea that decision-makers who “think a goverance structure will yield total human disempowerment” will “do something else” also seems quite implausible. Such decision-makers will likely struggle to retain power. Decision-makers who prioritize their own “power” (and feel empowered even as they hand off increasing decision-making to AI) and their immediate political survival above all else will be empowered.
Another features of the future which seems likely and can already be witnessed beginning is the gradual emergence and ascendance of pro-AI-takeover and pro-arms-race ideologies, which endorse the more competitive moves of rapidly handing off power to AI systems in insufficiently cooperative ways.
This thought experiment is described in ARCHES FYI. https://acritch.com/papers/arches.pdf
I think it’s a bit sad that this comment is being so well-received—it’s just some opinions without arguments from someone who hasn’t read the paper in detail.
There are 2 senses in which I agree that we don’t need full on “capital V value alignment”:
We can build things that aren’t utility maximizers (e.g. consider the humble MNIST classifier)
There are some utility functions that aren’t quite right, but are still safe enough to optimize in practice (e.g. see “Value Alignment Verification”, but see also, e.g. “Defining and Characterizing Reward Hacking” for negative results)
But also:
Some amount of alignment is probably necessary in order to build safe agenty things (the more agenty, the higher the bar for alignment, since you start to increasingly encounter perverse instatiation-type concerns—CAVEAT: agency is not a unidimensional quantity, cf: “Harms from Increasingly Agentic Algorithmic Systems”).
Note that my statement was about the relative requirements for alignment in text domains vs. real-world. I don’t really see how your arguments are relevant to this question.
Concretely, in domains with vision, we should probably be significantly more worried that an AI system learns something more like an adversarial “hack” on it’s values leading to behavior that significantly diverges from things humans would endorse.
OTMH, I think my concern here is less:
“The AI’s values don’t generalize well outside of the text domain (e.g. to a humanoid robot)”
and more:“The AI’s values must be much more aligned in order to be safe outside the text domain”
I.e. if we model an AI and a human as having fixed utility functions over the same accurate world model, then the same AI might be safe as a chatbot, but not as a robot.
This would be because the richer domain / interface of the robot creates many more opportunities to “exploit” whatever discrepancies exist between AI and human values in ways that actually lead to perverse instantiation.
Two things that strike me:
The claim that “There are three kinds of genies: Genies to whom you can safely say ‘I wish for you to do what I should wish for’; genies for which no wish is safe; and genies that aren’t very powerful or intelligent.” only seems true under a very conservative notion of what it means for a wish to be “safe” (which may be appropriate in some cases). It’s a very black-and-white account—certainly there ought to be a continuum of genies with different safety/performance trade-offs resulting from their varying capabilities and alignment properties.
The final 3 paragraphs of the linked post on Artificial Addition seem to suggest that deep learning-style approaches to teaching AI systems arithmetic are not promising. I also recall that EY and others thought deep learning wouldn’t work for capabilities, either. The argument that deep learning won’t work for capabilities has mostly been falsified. It seems like the same argument was being used to illustrate a core alignment difficulty in this post, but it’s not entirely clear to me.
This comment made me reflect on what fragility of values means.
To me this point was always most salient when thinking about embodied agents, which may need to reliably recognize something like “people” in its environment (in order to instantiate human values like “try not to hurt people”) even as the world changes radically with the introduction of various forms of transhumanism.
I guess it’s not clear to me how much progress we make towards that with a system that can do a very good job with human values when restricted to the text domain. Plausibly we just translate everything into text and are good to go? It makes me wonder where we’re at with adversarial robustness of vision-language models, e.g.
OK, so it’s not really just your results? You are aggregating across these studies (and presumably ones of “Westerners” as well)? I do wonder how directly comparable things are… Did you make an effort to translate a study or questions from studies, or are the questions just independently conceived and formulated?
No, I was only responding to the the first part.
Not necessarily fooling it, just keeping it ignorant. I think such schemes can plausibly scale to very high levels of capabilities, perhaps indefinitely, since intelligence doesn’t give one the ability to create information from thin air...
This is a super interesting and important problem, IMO. I believe it already has significant real world practical consequences, e.g. powerful people find it difficult to avoid being surrounded by sychophants: even if they really don’t want to be, that’s just an extra constraint for the sychophants to satisfy (“don’t come across as sychophantic”)! I am inclined to agree that avoiding power differentials is the only way to really avoid these perverse outcomes in practice, and I think this is a good argument in favor of doing so.
--------------------------------------
This is also quite related to an (old, unpublished) work I did with Jonathan Binas on “bounded empowerment”. I’ve invited you to the Overleaf (it needs to clean-up, but I’ve also asked Jonathan about putting it on arXiv).
To summarize: Let’s consider this in the case of a superhuman AI, R, and a human H. The basic idea of that work is that R should try and “empower” H, and that (unlike in previous works on empowerment), there are two ways of doing this:
1) change the state of the world (as in previous works)
2) inform H so they know how to make use of the options available to them to achieve various ends (novel!)
If R has a perfect model of H and the world, then you can just compute how to effectively do these things (it’s wildly intractable, ofc). I think this would still often look “patronizing” in practice, and/or maybe just lead to totally wild behaviors (hard to predict this sort of stuff...), but it might be a useful conceptual “lead”.
Random thought OTMH: Something which might make it less “patronizing” is if H were to have well-defined “meta-preferences” about how such interactions should work that R could aim to respect.
What makes you say this: “However, our results suggest that students are broadly less concerned about the risks of AI than people in the United States and Europe”?
This activation function was introduced in one of my papers from 10 years ago ;)
See Figure 2 of https://arxiv.org/abs/1402.3337
Really interesting point!
I introduced this term in my slides that included “paperweight” as an example of an “AI system” that maximizes safety.
I sort of still think it’s an OK term, but I’m sure I will keep thinking about this going forward and hope we can arrive at an even better term.
You could try to do tests on data that is far enough from the training distribution that it won’t generalize in a simple immitative way there, and you could do tests to try and confirm that you are far enough off distribution. For instance, perhaps using a carefully chosen invented language would work.
I don’t disagree… in this case you don’t get agents for a long time; someone else does though.
I meant “other training schemes” to encompass things like scaffolding that deliberately engineers agents using LLMs as components, although I acknowledge they are not literally “training” and more like “engineering”.
I would look at the main FATE conferences as well, which I view as being: FAccT, AIES, EEAMO.
(more than downward—not more than previous surveys)